JP3457578B2 - Speech recognition apparatus and method using speech synthesis - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device and a voice recognition method using voice synthesis.

[0002]

2. Description of the Related Art A conventional speech recognition apparatus based on phoneme-based matching will be described with reference to FIG.

In FIG. 6, a "recognition dictionary creating section" creates a "recognition word dictionary" by extracting information necessary for the recognition process from the "registered word character string" to be recognized.

The "input voice analysis section" extracts the "feature pattern" of the voice input by the speaker from the microphone. The "recognition matching unit" performs recognition processing by phoneme-based matching using the "recognition word dictionary" and "standard pattern" created above, and outputs recognition result candidates that are "recognition results".

However, the conventional speech recognition device based on phoneme-based matching has the following problems. That is, when words that are acoustically similar to each other are registered when the words are registered, there are many common phonemes that compose them. Therefore, the Markov model patterns used in the matching process become similar to each other, and it becomes difficult to identify both words at the time of recognition.

Therefore, the rate of erroneous recognition increases, and it becomes necessary to ask the speaker to utter again in order to increase the accuracy of the recognition result.

Japanese Patent Laid-Open No. 9-6387 discloses an example of a technique that improves this conventional example.

This conventional example is an apparatus for the purpose of providing a speech recognition apparatus having an excellent ability to discriminate between similar words. As shown in FIG. 7, the “word voice cutout unit” cuts out a word voice from the voice input by the microphone that is the voice input unit, and the “feature extraction unit” extracts the feature data.

The "state number estimating section" estimates the number of states for word speech when modeling from feature data by a Markov model. The “similar word determination unit” determines whether or not a word similar to the word voice to be newly registered has already been registered.

The "state number adding unit" increases the estimated number of states, and the "learning unit" applies the feature data to the word model to obtain Markov model parameters.

The "matching determination unit" performs likelihood calculation for each word model, determines a recognition candidate, and outputs a recognition result from the "determination result output unit". The "matching determination unit" uses the "voice dictionary file" composed of the learned Markov model parameters in the matching process.

[0012]

However, the voice recognition device improved as described above has the following problems.

Since it is necessary to learn the Markov model parameters based on the uttered voice when registering the similar words, it is not easy to register the recognition target words.

Since it is necessary to register by utterance,
In order to improve the recognition rate of the unspecified speaker, it becomes necessary to collect voicing voices from many speakers.

An object of the present invention is to expand the difference between similar words, and when similar words are registered as recognition targets,
An object of the present invention is to provide a voice recognition device and method capable of reducing recognition error.

[0016]

A speech recognition apparatus according to the present invention comprises a similar word dictionary section in which words having similar phonemes are stored, and a recognition word dictionary section in which recognition words used in normal recognition processing are stored. A synthesized voice feature pattern portion for storing respective feature patterns of synthesized voice formed from data of the similar word dictionary portion, voice input means, and the voice
An input speech analysis section for forming a feature pattern of the input input speech by the input means, the recognition matching unit for outputting a recognition result candidates from the data of the recognition word dictionary, wherein the pattern of the input speech, the recognition result candidate Determination means for determining whether or not a similar word exists, and a feature pattern of the input voice when the determination result determines that a similar word exists and a similar word stored in the similar word dictionary unit A first similar word matching unit that performs a matching process for each and outputs a first similar word recognition result; and, when the determination result determines that there is a similar word ,
Of the characteristic pattern and the data of the synthetic speech characteristic pattern
A second similar word matching unit that performs a matching process and outputs a second similar word recognition result for each of the similar words is compared with the first similar word recognition result and the second similar word recognition result to determine the accuracy. It is characterized by having a similar word recognition result comparison unit that outputs a high word as a recognition result.

The speech recognition method of the present invention comprises the steps of storing words having similar constituent phonemes in a similar word dictionary section, storing a recognized word used in normal recognition processing in the recognized word dictionary section, The step of storing the respective characteristic patterns of the synthesized speech formed from the data of the word dictionary section in the synthesized speech characteristic pattern section, the speech input step, and the input forming the characteristic pattern of the input speech input by the speech input step. A voice analysis step, and
A recognition matching step of outputting a recognition result candidate from the feature pattern of the input voice and the data of the recognition word dictionary;
A determination step of determining whether or not a similar word exists in the recognition result candidate; and the input voice feature pattern and the similarity when the determination result determines that a similar word exists.
Matching process for each similar word stored in the word dictionary
A first similar word matching step of performing a processing to output a first similar word recognition result; and a characteristic pattern of the input voice when the determination result determines that a similar word exists.
Matching processing with the data of the synthetic voice feature pattern
And a second similar word matching step for outputting a second similar word recognition result for each similar word, and the first similar word recognition result and the second similar word recognition result are compared, and a word with high accuracy is determined. And a similar word recognition result comparing step of outputting as a recognition result.

[0018]

BEST MODE FOR CARRYING OUT THE INVENTION Next, a first embodiment of the present invention will be described with reference to the drawings.

In FIG. 1, a "recognition dictionary creating section" creates a recognition word dictionary 2 by extracting information necessary for a recognition process from a registered word character string input as a recognition target, and at the same time, a constituent phoneme is generated. Similar words are extracted to create a similar word dictionary 1.

The "warning output section" presents the extracted similar word to the speaker to call attention to the speaker.

The "input voice analysis section" extracts the characteristic pattern from the voice input from the microphone.

The "recognition matching unit 3" performs matching processing based on the standard pattern previously learned for each phoneme, the recognition word dictionary, and the above-mentioned characteristic pattern, and the word having the highest similarity to the input speech. Is output as a recognition result candidate.

In order to obtain more detailed information, the "recognition result candidate determination unit" is performed for the output recognition result candidates including similar words and having a small difference in recognition likelihood between the candidates. It is determined whether or not to perform recognition processing. If the recognition result candidates do not include similar words, or the recognition likelihood between candidates is high, that is, if the probability that the first candidate is the word uttered is high, the next recognition process is not performed and Output as a recognition result.

The "synthesized speech output unit" outputs a synthesized speech waveform from the input text of similar words.

The "synthesized speech analysis section" extracts the characteristic pattern from the synthesized speech waveform.

The "similar word recognition matching section 1" performs matching processing for similar words based on the input feature pattern of the voice, and outputs a similar word recognition result 1.

The "similar word recognition matching unit 2" performs a matching process for similar words based on the characteristic pattern extracted from the synthetic speech, and outputs a similar word recognition result 2.

The "similar word recognition result comparing section" determines and outputs a final recognition result based on the similar word recognition result 1 and the similar word recognition result 2.

As described above, according to the present invention, the recognition result output by the voice recognition process using the synthetic voice waveform as an input and the recognition result output by the voice recognition process for the similar words extracted from the recognition target words. Since the recognition result determination is performed together with the result, it is possible to improve the recognition performance in the recognition target words having similar constituent phonemes.

Referring to FIG. 1, there is shown a voice recognition apparatus as an embodiment of the present invention. In the figure, this embodiment is similar to the recognition dictionary creation unit, the warning output unit, the input voice analysis unit, the recognition matching unit, the recognition result candidate determination unit, the synthetic voice output unit, the synthetic voice analysis unit, and the similar word recognition matching unit 1. The word recognition matching unit 2 and the similar word recognition result comparison unit are included.

The configuration will be described in detail with reference to FIG.

The operation of the recognition dictionary creating section of FIG. 1 is shown in FIG.
This will be described using the flowchart shown in.

First, the input registered word character string is converted into a phoneme string forming each word (STEP 10).

Next, based on the converted phoneme sequence, words having similar constituent phonemes are detected from the registered word group (S
TEP11). At this time, for example, by setting a threshold value for the value indicating the ratio of the constituent phoneme strings that match between words to the entire word, it is possible to determine whether or not they are similar.

Next, in STEP 11, it is judged whether or not a similar word is detected. If it is detected, the process proceeds to STEP 13, and if it is not detected, the process proceeds to STEP 14 (STEP 12).

In STEP 13, a similar word dictionary is created from the detected similar words. As the information stored in the similar word dictionary, in addition to the phoneme sequence that forms a word, for example, a mixture number that indicates a Gaussian distribution for each phoneme that forms a similar word is expanded from a value in a normal recognition process or a similar word There are notations (things that show the accent information when speaking).

At STEP 14, a recognition word dictionary used in normal recognition processing is created. For example, the phoneme string information that constitutes a word is stored as a recognized word dictionary.

Next, the operation in the process of inputting voice from the microphone of FIG. 1 and determining the recognition result candidate will be described with reference to the flowchart shown in FIG.

First, it is determined whether a similar word is detected from the words to be recognized (STEP 20).
Proceed to STEP21, and if not detected, proceed to STEP22.

At STEP 21, the warning output unit presents a list of similar words to the speaker to call attention to the utterance.

In STEP 22, the voice input from the microphone is analyzed and the characteristic pattern is output.

At this time, the characteristic pattern can be obtained by performing the mel cepstrum analysis described in “Voice Recognition” by Sei Imai, Kyoritsu Shuppan Co., Ltd. (hereinafter referred to as Reference 1).

Next, a matching process between the obtained characteristic pattern and standard pattern is performed (STEP 23). For example, the distance between the feature pattern and the standard pattern can be calculated by the method using the DP matching method and the HMM described in Document 1, and the cumulative distance of each word to be recognized can be calculated.

From the words to be recognized, the word with high processing likelihood obtained in STEP 23 is output as a recognition result candidate (STEP 2
Four).

Next, it is determined whether the output recognition result candidates include those detected as similar words (S
TEP25). If it is included, go to STEP 26. If not, go to STEP 26.
Go to STEP31.

At STEP 26, the accuracy of the recognition result candidate is determined from the likelihood of the similar word included in the recognition result candidate. For example, the likelihood difference between similar word candidates can be used as the criterion for the determination. The threshold value of the likelihood difference used at the time of determination can be determined in advance by evaluating the system, for example.

Next, it is determined whether or not the similar word recognition process is performed based on the accuracy of the recognition result candidate (STEP 27). If you do
Proceed to STEP28. If not, proceed to STEP31.

At STEP 28, similar word recognition processing 1 is performed.
Details of the processing here will be described separately with reference to FIG.

At STEP 29, similar word recognition processing 2 is performed.
Details of the processing here will be described separately with reference to FIG.

At STEP 30, the similar word recognition results obtained at STEP 28 and STEP 29 are compared and examined, and the word judged to have high accuracy is output as the final recognition result.

Next, referring to FIG. 4, there is shown a flowchart of the similar word recognition matching process 1 in one embodiment of the present invention.

First, the characteristic pattern obtained in STEP 22 in FIG. 3 is input, and matching processing is performed on similar words (STEP 100). At this time, a more accurate result can be obtained by performing the matching process with the mixture number added in STEP 13 in FIG. 2 and showing the Gaussian distribution for the constituent phonemes of the similar word.

Next, the recognition result for each similar word is output as the similar word recognition result 1 (STEP 101).

Next, referring to FIG. 5, there is shown a flowchart of the similar word recognition matching process 2 in the embodiment of the present invention.

First, the similar word character string is input to the speech synthesis processing, and each synthesized speech is output (STEP 200).
Here, as a method of outputting a synthetic voice from a word character string, for example, a text is syntactically analyzed, which is described in “Voice Processing and DSP” by Keigaku Shuppan, Yasuhiko Arai and Masami Osaki (hereinafter referred to as Reference 2). After that, there is a method of editing the phoneme of the obtained phoneme.

Next, the output synthetic speech is analyzed and each characteristic pattern is output (STEP 201). The analysis process at this time may be the same as that of STEP22 in FIG.

Next, a matching process is performed between the characteristic pattern of the synthesized voice obtained in STEP 201 and the characteristic pattern of the input voice obtained in STEP 22 in FIG. 3 to obtain the distance value between them (STEP 202).

Next, the distance value between the synthesized voice and the voiced voice for each similar word obtained in STEP 202 is output as a similar word recognition result 2 (STEP 203).

As another embodiment of the present invention, the basic structure is as described above, but the "similar word recognition matching unit 1" and the "similar word recognition matching unit 2" are combined into one, and "similar word recognition" is performed. It may be a “matching unit”.

In this case, the "input speech analysis section" extracts an extended characteristic pattern in which the number of parameters is extended in addition to the characteristic pattern when a similar word is detected. The number of parameters can be expanded by, for example, a method of increasing the number of dimensions of the mel cepstrum extracted in the processing of STEP22 in FIG. Similarly, the "synthetic speech analysis unit" also extracts an extended characteristic pattern for the input synthetic speech waveform. The "similar word recognition matching unit" performs a matching process between the extended feature pattern of the input voice and the extended feature pattern for the synthetic speech waveform to obtain a distance value, and the "similar word recognition result comparison unit" calculates the degree of similarity between them. The recognition result is determined by. The matching processing here may be performed by the method using the DP matching method and the HMM described in Document 1 as in the normal matching processing.

Further, as another embodiment, the basic structure thereof is as described above, but as the means for obtaining the accent information of the similar word among the information stored as the similar word dictionary, the embodiment described above is used. It is also possible to use a combination of the kana of the word and the accent information, instead of the notation of the word used in the inside. In this case, the synthesized voice output unit can omit the process of analyzing the notation and converting it into pronunciation information corresponding to the output voice waveform.

[0062]

According to the present invention, as described above,
The present invention has the following effects.

The first effect is to extract words having similar phoneme strings from the recognition target words and perform matching processing using more detailed feature patterns for them.
It is to increase the difference between similar words and reduce the recognition error for those utterances when similar words are registered as recognition targets.

The second effect is that it is possible to automatically create a characteristic pattern that is uniquely determined by using voice synthesis capable of outputting a voice in a fixed pattern. In particular, a voice recognition function and a voice synthesis function are installed. In the system, it is not necessary to perform new learning when expanding the characteristic parameter, and the area for storing the expanded characteristic parameter is not necessary.

The third effect is that by clearly indicating to the speaker words similar to the phoneme string extracted in advance, it is possible to encourage careful utterance at the time of utterance. Immediately after the correction circuit corrects the retention failure, it becomes the target immediately.

[Brief description of drawings]

FIG. 1 is a diagram of a system showing a first embodiment of the present invention.

FIG. 2 is a diagram showing a flow of creating a recognized word dictionary shown in FIG.

FIG. 3 is a diagram showing a flow until the recognition result output shown in FIG.

FIG. 4 is a diagram showing a flow of a similar word recognition process shown in FIG. 1.

5 is a diagram showing a flow of a similar word recognition process 2 shown in FIG.

FIG. 6 is a diagram showing a first conventional example.

FIG. 7 is a diagram showing a second conventional example.

[Explanation of symbols]

1 Similar word dictionary 2 recognition word dictionary 3 Recognition matching section 4 Similar word recognition matching unit 1 5 Similar word recognition matching unit 2 6 Similar word recognition result comparison unit

Front page continuation (58) Fields surveyed (Int.Cl. ⁷ , DB name) G10L 15/06 G10L 15/22

Claims (6)

(57) [Claims] 1. A similar word dictionary unit in which words having similar constituent phonemes are stored, a recognition word dictionary unit in which recognition words used in normal recognition processing are stored, and data of the similar word dictionary unit. and synthesized speech feature pattern unit for storing each of the feature patterns of the synthesized speech that, an audio input unit, an input speech analysis section for forming a feature pattern of the input input speech by said speech input means, said input speech < A recognition matching unit that outputs a recognition result candidate from the characteristic pattern and the data of the recognition word dictionary; a determination unit that determines whether or not a similar word exists in the recognition result candidate;
When the determination result determines that there is a similar word,
The characteristic pattern of the input voice and the similar word stored in the dictionary section.
A first similar word matching unit that performs matching processing for each similar word that is output and outputs a first similar word recognition result, and the determination result is that a similar word exists
The characteristic pattern of the input voice and the synthetic voice feature pattern
A second similar word matching unit that performs a matching process with the data of the second similar word and outputs a second similar word recognition result for each similar word; and the first similar word recognition result and the second similar word recognition result. A speech recognition apparatus using speech synthesis, comprising: a similar word recognition result comparison unit that outputs a highly accurate word as a recognition result by comparison. 2. The voice recognition device using voice synthesis according to claim 1, further comprising a warning output unit for generating a caution when the similar word is detected when the similar word is detected. 3. The similar word dictionary unit converts an input registered word character string into a phoneme string forming each word, and stores words having similar phonemes as a similar word. Recognition device using the above speech synthesis. 4. A step of storing words having similar phonemes in a similar word dictionary section, a step of storing a recognition word used for normal recognition processing in the recognition word dictionary section, and data from the similar word dictionary section. it synthesized speech to be formed
Storing the feature pattern Les synthetic speech feature pattern portion, a voice input step, the speech input step
An input speech analysis step of forming a feature pattern of the input input speech by flop, a recognition matching step of outputting the recognition result candidates from data of the feature pattern and the recognition word dictionary of the input speech, similar to the recognition result candidates whether a determination step of determining a word is present, the input when the judgment result is judged as similar word exists
Stored in the similar voice dictionary and the force voice feature pattern
A first similar word matching step in which a matching process is performed for each similar word that is present and a first similar word recognition result is output; and when the determination result determines that a similar word exists
The characteristic pattern of the input voice and the synthetic voice feature pattern
Second matching for each similar word
Second similar word matching step for outputting the similar word recognition result of the above and the similar word recognition for outputting a highly accurate word as a recognition result by comparing the first similar word recognition result and the second similar word recognition result. And a result comparing step. 5. The voice recognition method according to claim 4, further comprising a warning output step of generating a caution when the similar word is detected when the similar word is detected. 6. The similar word dictionary unit converts an input registered word character string into a phoneme string forming each word and stores words having similar phonemes as a similar word. Voice recognition method.